Speaker: Thais Fonseca (University of Warwick)
Title: Dynamical Bayesian Networks for non-stationary discrete data
Microsoft Teams Meeting (link to appear)
In this work, a dynamical model is considered for time series of multivariate counts when an underlying BN is assumed. Based on conditional independence statements a BN provides a flexible tool to model multivariate data. Although there are several approaches available for time-varying networks of continuous data, few attempts consider temporal dependence for discrete BN. Independence is usually assumed due to computational feasibility rather than good fit to the data. However, in many applied settings it is unrealistic to assume that the conditional probability tables are stationary over time. In this work, a new class of state-space models is considered to allow for conditional probability tables to evolve smoothly over time. The proposal is based on the Dirichlet evolutional process which is analytically available making the computation feasible even for large datasets and long temporal windows. An application to multivariate extremes illustrates the effectiveness of our proposal. In particular, the probability of crossing thresholds are modelled for a vector of pollutants over many years, changes of regime due to interventions are identified and predictive performance is evaluated.